GenDexGrasp is a novel, hand-agnostic grasping algorithm that addresses the challenges of generalizable robotic grasping. Trained on the large-scale MultiDex dataset, it efficiently generates diverse and successful grasps for various multi-fingered robotic hands, outperforming previous methods in success rate, inference speed, and diversity.
In recent years, the pursuit of general artificial intelligence has intensified, but challenges remain. One major hurdle is developing Artificial Social Intelligence (ASI), which encompasses perceiving social events, inferring goals and intentions, and facilitating interactions. This complex ability sets humans apart from primates and remains difficult to implement in AI systems.
Recently, our robotics team published a paper in IJCV 2022 titled “Scene Reconstruction with Functional Objects for Robot Autonomy.” They introduced an innovative problem of scene reconstruction and representation, offering essential information for robotic autonomy while enabling interactive virtual scenes that closely mimic real-world scenarios for simulation testing.
A team led by Zhu Songchun, professor at the Institute of AI of Peking University, has built a computing framework for robots to understand human values in real time, enabling robots and human users to complete a series of complex human-robot collaboration tasks through real-time communication.